Hyperparameters are configuration variables controlling the behavior of machine learning algorithms. They are ubiquitous in machine learning and artificial intelligence and the choice of their values determine the effectiveness of systems based on these technologies. Manual hyperparameter search is often unsatisfactory and becomes unfeasible when the number of hyperparameters is large. Automating the search is an important step towards automating machine learning, freeing researchers and practitioners alike from the burden of finding a good set of hyperparameters by trial and error. In this survey, we present a unified treatment of hyperparameter optimization, providing the reader with examples and insights into the state-of-the-art. We cover the main families of techniques to automate hyperparameter search, often referred to as hyperparameter optimization or tuning, including random and quasi-random search, bandit-, model- and gradient- based approaches. We further discuss extensions, including online, constrained, and multi-objective formulations, touch upon connections with other fields such as meta-learning and neural architecture search, and conclude with open questions and future research directions.

We report about the application of state-of-the-art deep learning techniques to the automatic and interpretable assignment of ICD-O3 topography and morphology codes to free-text cancer reports. We present results on a large dataset (more than 80 000 labeled and 1 500 000 unlabeled anonymized reports written in Italian and collected from hospitals in Tuscany over more than a decade) and with a large number of classes (134 morphological classes and 61 topographical classes). We compare alternative architectures in terms of prediction accuracy and interpretability and show that our best model achieves a multiclass accuracy of 90.3% on topography site assignment and 84.8% on morphology type assignment. We found that in this context hierarchical models are not better than flat models and that an element-wise maximum aggregator is slightly better than attentive models on site classification. Moreover, the maximum aggregator offers a way to interpret the classification process.

We study the problem of fitting task-specific learning rate schedules from the perspective of hyperparameter optimization. This allows us to explicitly search for schedules that achieve good generalization. We describe the structure of the gradient of a validation error w.r.t. the learning rate, the hypergradient, and based on this we introduce a novel online algorithm. Research in this direction is vast (see Hutter et al. (2019) for an overview) and includes model-based (Snoek et al., 2012; Hutter et al., 2015), model-free (Bergstra & Bengio, 2012; Hansen, 2016), and gradient-based (Domke, 2012; Maclaurin et al., 2015) approaches. Problem (1) can be in principle solved by any HPO technique.

We introduce an extension of the multi-instance learning problem where examples are organized as nested bags of instances (e.g., a document could be represented as a bag of sentences, which in turn are bags of words). This framework can be useful in various scenarios, such as text and image classification, but also supervised learning over graphs. As a further advantage, multi-multi instance learning enables a particular way of interpreting predictions and the decision function. Our approach is based on a special neural network layer, called bag-layer, whose units aggregate bags of inputs of arbitrary size. We prove theoretically that the associated class of functions contains all Boolean functions over sets of sets of instances and we provide empirical evidence that functions of this kind can be actually learned on semi-synthetic datasets. We finally present experiments on text classification and on citation graphs and social graph data, showing that our model obtains competitive results with respect to other approaches such as convolutional networks on graphs.

We introduce a framework based on bilevel programming that unifies gradient-based hyperparameter optimization and meta-learning. We show that an approximate version of the bilevel problem can be solved by taking into explicit account the optimization dynamics for the inner objective. Depending on the specific setting, the outer variables take either the meaning of hyperparameters in a supervised learning problem or parameters of a meta-learner. We provide sufficient conditions under which solutions of the approximate problem converge to those of the exact problem. We instantiate our approach for meta-learning in the case of deep learning where representation layers are treated as hyperparameters shared across a set of training episodes. In experiments, we confirm our theoretical findings, present encouraging results for few-shot learning and contrast the bilevel approach against classical approaches for learning-to-learn.

In (Franceschi et al., 2018) we proposed a unified mathematical framework, grounded on bilevel programming, that encompasses gradient-based hyperparameter optimization and meta-learning. We formulated an approximate version of the problem where the inner objective is solved iteratively, and gave sufficient conditions ensuring convergence to the exact problem. In this work we show how to optimize learning rates, automatically weight the loss of single examples and learn hyper-representations with Far-HO, a software package based on the popular deep learning framework TensorFlow that allows to seamlessly tackle both HO and ML problems.

We consider a class of a nested optimization problems involving inner and outer objectives. We observe that by taking into explicit account the optimization dynamics for the inner objective it is possible to derive a general framework that unifies gradient-based hyperparameter optimization and meta-learning (or learning-to-learn). Depending on the specific setting, the variables of the outer objective take either the meaning of hyperparameters in a supervised learning problem or parameters of a meta-learner. We show that some recently proposed methods in the latter setting can be instantiated in our framework and tackled with the same gradient-based algorithms. Finally, we discuss possible design patterns for learning-to-learn and present encouraging preliminary experiments for few-shot learning.

We study two procedures (reverse-mode and forward-mode) for computing the gradient of the validation error with respect to the hyperparameters of any iterative learning algorithm such as stochastic gradient descent. These procedures mirror two methods of computing gradients for recurrent neural networks and have different trade-offs in terms of running time and space requirements. Our formulation of the reverse-mode procedure is linked to previous work by Maclaurin et al. [2015] but does not require reversible dynamics. The forward-mode procedure is suitable for real-time hyperparameter updates, which may significantly speed up hyperparameter optimization on large datasets. We present experiments on data cleaning and on learning task interactions. We also present one large-scale experiment where the use of previous gradient-based methods would be prohibitive.

We introduce an architecture based on deep hierarchical decompositions to learn effective representations of large graphs. Our framework extends classic R-decompositions used in kernel methods, enabling nested "part-of-part" relations. Unlike recursive neural networks, which unroll a template on input graphs directly, we unroll a neural network template over the decomposition hierarchy, allowing us to deal with the high degree variability that typically characterize social network graphs. Deep hierarchical decompositions are also amenable to domain compression, a technique that reduces both space and time complexity by exploiting symmetries. We show empirically that our approach is competitive with current state-of-the-art graph classification methods, particularly when dealing with social network datasets.

We introduce kLog, a novel approach to statistical relational learning. Unlike standard approaches, kLog does not represent a probability distribution directly. It is rather a language to perform kernel-based learning on expressive logical and relational representations. kLog allows users to specify learning problems declaratively. It builds on simple but powerful concepts: learning from interpretations, entity/relationship data modeling, logic programming, and deductive databases. Access by the kernel to the rich representation is mediated by a technique we call graphicalization: the relational representation is first transformed into a graph --- in particular, a grounded entity/relationship diagram. Subsequently, a choice of graph kernel defines the feature space. kLog supports mixed numerical and symbolic data, as well as background knowledge in the form of Prolog or Datalog programs as in inductive logic programming systems. The kLog framework can be applied to tackle the same range of tasks that has made statistical relational learning so popular, including classification, regression, multitask learning, and collective classification. We also report about empirical comparisons, showing that kLog can be either more accurate, or much faster at the same level of accuracy, than Tilde and Alchemy. kLog is GPLv3 licensed and is available at http://klog.dinfo.unifi.it along with tutorials.